Skip to main content

Agentic Runtime Blueprint

This is the practical blueprint for how Nova behaves as an agent.

What "agentic" means in Nova​

Nova can understand natural text, choose an action path, and execute it safely without requiring the user to learn command syntax.

Runtime contract​

For every instruction, Nova returns one of:

  • ok: completed safely
  • needs_approval: proposal prepared and waiting for user confirmation
  • failed: no unsafe mutation happened; user gets a clear reason

Simple sequence​

Message
  -> classify intent
  -> build structured plan
  -> validate plan
  -> run policy checks
  -> execute or request approval
  -> store audit + return result

Clarification behavior​

Nova asks for clarification only when needed:

  • missing date/time/person
  • ambiguous target (for example multiple matching meetings)
  • low confidence after one internal repair attempt

Confirmation behavior​

For messages like "yes" or "do it":

  • one pending approval -> execute
  • none pending -> explain there is nothing pending
  • multiple pending -> ask user to pick one

Calendar behavior (user-facing)​

Read requests:

  • "what do I have tomorrow?"
  • "what did I have yesterday?"
  • "what do I have on March 15?"

Write requests:

  • create, reschedule, cancel
  • always proposal-first for mutating actions

Timezone fallback:

  1. user profile timezone
  2. chat/runtime timezone context
  3. workspace timezone
  4. UTC

Reliability guarantees​

  • no unbounded scans in critical paths
  • replay-protected confirmations
  • idempotent mutation execution
  • explicit partial status for mixed flows